Monday, November 26, 2007

AMD's 65nm Process Hurts AMD

In our blogosphere neighborhood Roborat64 addressed the topic of the involution of AMD's product line. I would like to complement his words with a bit of updated concrete evidence.

In "New > OLD" [ Note: I tend to shorten the names of articles if they are properly linked to, this article originally is named "The Fundamental Law of Progess [ sic ] NEW > OLD" ], Roborat64 says: "the fundamental law of progress [...] demands all NEW PRODUCTS [TO] BE BETTER THAN OLD PRODUCTS", but "it is increasingly alarming how AMD seemed to be moving backwards". "The introduction of slower 65nm CPUs and now K10's inferior performance to K8 are just some of the bad habits AMD seemed to be developing".

If we rewind one year to the introduction of 65nm K8, we will see that, from the point of view of the customer, AMD offered worse new products: Processors with some increased L2 cache latencies, that is, an slightly smaller IPC, and clocks not as fast as the fastest 90nm products. At that time, some people thought that AMD emphasized low production costs over higher premiums because AMD was production bounded, especially due to the possibly great increase in demand coming from Dell [ Many other posts contain abundant references as to the catastrophic consequences of this strategy ], and then the optimists assumed that there was no problem with AMD's 65nm process, the company was going to be able to quickly jack up clock speeds in the new process. But that didn't happen. The optimist then insisted on the superior power efficiency of the 65nm, but unfortunately the market cares very little about this parameter and they don't want to accept this fact.

I predicted serious problems with the speeds of the 65nm as soon as I saw AMD's explanation for the increased L2 latencies in 65nm K8: Supposedly, the architecture increased the latencies to leave room for future L2 expansions. Taking into account that the L2 of Brisbane already was half the capacity of the high end 90nm, 512MB versus 1MB per core, and that AMD "could cross that bridge when it got there" for a subject as simple as L2 parameters, there was no other way to interpret AMD's explanations as bullshit, that the company was forced to cut corners because it could not do a good shrink. Taking into account that the 65nm was introduced later and in smaller quantities than expected, I hypothesized that the problems were serious. To confirm this appreciation, AMD has launched ever more disappointing processors, and contrary to any tradition, the products of highest clocks are all old process!.

We must visit AMD's official pages to confirm this highly unusual situation, go to and select both "90nm" and "65nm" in the "Manufacturing Tech (CMOS)" pull down list:

90nm top of the line: 6400+: 3.2 GHz with 2 MB L2, 1.35/1.40 V, 125 Watts.

65nm top of the line: 5200+: 2.7 GHz with 1 MB L2 (half the size), 65 Watts (half consumption, or roughly 61% power consumption accounting for the speed difference, much better, but then again, what does it matter that it consumes less power if things are slower?)

It may be argued that the top of the line in 65nm product is the Phenom 9600, the quadcore @ 2.3 GHz. Too slow, that is what I say. It is so slow, that it has been proven slower in many workloads than AMD's own dual cores... [ link may be found here ]

A very similar thing happens in Opterons:
No 65nm dual core Opteron, measly 2.0 GHz top of the line Quadcore, the 2350 and 8350 [ link ], and a top clock of 3.2 GHz for the 90nm top of the line dual cores, the 8224 SE and 2224 SE [ link ]

AMD's own price lists demonstrate the horrible catastrophe the company is in:

Family Model Process Freq QC/DC Price Price per core
Opteron 2350 65nm 2 QC $389 $97.25

8350 65nm 2 QC $1,019 $254.75

2224 SE 90nm 3.2 DC $873 $436.50

8224 SE 90nm 3.2 DC $2,149 $1,074.50
Athlon 64 X2 6400+ 90nm 3.2 DC $220 $110.00

5200+ 65nm 2.7 DC $125 $62.50
Athlon 64 FX FX 74 90nm 3 DC $300 $150.00

An interesting first observation is that the price per core of 90nm may be more than four times (!!!) the price per core of 65nm processors...

We know that 65nm processors take more than half the die area of 90nm processors because the shrink was to an equivalent node size of over 70nm, we must also expect the yields of the 65nm process to be smaller, and the binsplits to be worse. Furthermore, in the case of quadcores, we also know that the yields and binsplits are quadratically worse than dual cores. From the point of view of production costs, the 65nm are not even twice as cheap, but from the point of view of ASPs, they are way less than half the 90nm, it is clear from the table above that the 65nm is actually hurting AMD! because it is more expensive to manufacture the 65nm product to obtain a smaller selling price.

It may be argued that this is a biased explanation because AMD chose the 65nm process for volume and 90nm for premiums, but this is illogical:
  1. We know that the market is saturated by entry-level crap (Intel Netbursts still being purged, Celerons, and even Yonah/Sossaman Core Duo/Solo incapable of 64 bits), thus the product moves only through steep discounts, thus if AMD would be able to, it would target the 65nm product for the upper segments.
  2. 65nm should have many inherent advantages regarding performance, by insisting on the semi-exhausted 90nm product for top of the line performance, AMD is mortgaging the future of the company.
AMD's words that they already shifted the focus to the 45nm process is a very laughable statement: It has been pretty much demonstrated that there is something serious on the 65nm process that hasn't been solved, why would it be any different at the 45nm process?. Perhaps it is something like that 65nm didn't allow for the same performance of the 90nm process due to inherent limitations, perhaps, the power consumption didn't go down in the same magnitude the heat dissipation problems of SOI went up, I don't know, I am just speculating, but what I do know is that there is a serious problem and AMD is denying its existence.

I feel as if I were beating a dead horse by saying that if the 65nm process didn't scale, it is of very little relevance that AMD will make immersion 45nm if it doesn't have the excellent advancements of High-Dielectric-Constant / Metal Gate transistors Intel is bringing to the market. It is also beating a dead horse to point out that K10 demonstrated that it is a dubious improvement on IPC due to the increased latencies in the L3 cache because Intel comes with Integrated Memory Controller/Point to Point interprocessor communication architectures with enlarged caches and higher clocks!. But what AMD and its cheerleaders are doing, of pegging new hopes on the 45nm process and the dual core K10 to come, is like putting the cart ahead of the horse.


Anonymous said...

Just curious - you previously said 65nm is just "Intel marketing"

...still think that?

"Core2 are dauntingly complex ยต-procs. This means that there is no architectural improvement coming; Intel decided to rely solely in shrinks to improve speeds and performance."

...Still think this?

"AMD does have the package to get to over 50% market share:

1. Good, proven management
2. Good, proven, exciting technology
3. The brand name has appreciated significantly, creating a loyal customer base
4. Technological leadership assures media exposure, free good publicity, public awareness, toghether with the enthusiasm of partners in the ecosystem
5. The ecosystem itself
6. Production Capacity!"

This too?

"Confirming that AMD is closer to the bleeding edge of technology than Intel, there is the potential of things such as Z-Ram (Zero Capacitor RAM with 5 times higher densities than regular six-transistor flip-flops). I don't know about the introduction rates for semiconductor technology, but it seems that very soon, like in the scale of one or two years, AMD will be able to include Z-Ram either as L2 or L3 cache memory."

That was ~1.5 years ago, how's that ZRAM working out? You were right about AMD be on the bleeding edge...the 65nm is bleeding them.

With all due respect looking over some of your old blogs, it appears as though you have minimal understanding of process technology and should not be commenting on it.

And here's a secret for you - EVEN if AMD implemented high K, immersion litho ADDS NOTHING TO PERFORMANCE - it is merely another technique to print the same size 45nm features (Intel chooses a 2 pass process on the critical steps instead). In the end the feature size is the same. Immersion is just another PR gimmick to make it seem like AMD is ahead, while the scary truth is Intel has far superior litho capabilities as they are able to EXTEND AND REUSE the older dry litho one generation further. Sorprisingly even with a 2 pass process, it is still cheaper than a single pass immersion process as the immersion tools are nearly 2X higher price, run at slower speeds and are less mature than existing dry litho technologies.

But then again what do I know, it seems like you have a good grasp on this whole technology thing.

Eddie said...

Hello there.

Thanks for your post, as a matter of fact I am going to post a full article regarding my screwups and your comment gives me a cue.

I would like your opinion about my explanations. Also, please try to post a name to refer to you, there is no need to sign up or anything, just to put something instead of Anonymous.